Muse Spark Open runs on two Apple M5 Ultra Mac Studios
Meta AI Research posted a blog entry titled “Muse Spark Open” on October 15, 2026, announcing that the 816‑billion‑parameter model is now available for download and can run on two Apple M5 Ultra Mac Studios.
Muse Spark Open is a multimodal model with 816 billion total parameters, of which 42 billion are active during inference, and it supports a one-million-token context window. It is positioned for coding and long-horizon reasoning tasks. The model is distributed under the “Muse Model License” with limited commercial terms, and the weights are hosted on platforms such as Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. In addition to direct download, the model is offered through API services including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Meta lists a suite of inference engines that can execute the model—SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth—each with optimizations for AMD, Apple, Intel and NVIDIA silicon.
The blog explains that the model runs on Apple’s M5 Ultra chip because its unified memory architecture lets the integrated GPU access up to 512 GB of RAM per Mac Studio. Two such machines together provide a terabyte of unified memory, sufficient to hold the entire model. The chip’s 1.2 TB/s memory bandwidth is described as fast enough to read the small slice of weights needed for each token, effectively letting the pair of Macs behave like a single supercomputer with a terabyte of memory. Meta demonstrates on-device agentic workflows such as modifying an iOS client app and reviewing DaVinci Resolve footage, all performed locally without cloud calls.
The system consists of the model weights stored in public repositories, the MLX framework on macOS, and the Apple M5 Ultra hardware with unified memory. When a user launches the model, MLX pulls the weights from the repository and maps them into the unified memory address space. Inference engines such as SGLang or vLLM then request the slice of weights needed to generate each token; the request travels over the CPU-GPU bus, and the GPU reads the slice directly from RAM at the advertised bandwidth. Because the memory is shared between CPU and GPU, the system avoids costly data copies between separate GPU memory and system memory, which would add latency. The same weights can also be loaded by other engines on AMD, Intel or NVIDIA GPUs, but only the Apple-specific path exploits the unified memory advantage.
The outcome—requiring two machines to hold the model—stems directly from the hardware constraint (512 GB per machine) and the design choice to rely on unified memory instead of distributing the model across many smaller GPUs. The trade-off is clear: Meta gains a tightly controlled, low-latency inference path that reinforces its ecosystem, while users bear a high upfront hardware cost. The licensing requirement further nudges developers toward self-hosting, which reduces price competition among cloud providers but also limits the model’s accessibility to organizations that can afford the Apple hardware.
An unresolved question remains: how effectively can Meta enforce the limited-commercial-terms license once the model can be executed on consumer-grade machines that users can freely copy and redistribute? The answer will determine whether the hardware barrier or the licensing model will be the dominant factor shaping the model’s adoption.